Self-supervised enhancement of stimulus-evoked brain response data

Stimulus-evoked brain response data has a notoriously low signal-to-noise ratio (SNR) and high inter-subject variability. Multiple techniques have been proposed to alleviate this problem, such as averaging, denoising source separation (DSS) and (multiway) canonical correlation analysis ((M)CCA), but all these methods have significant limitations. We propose a novel paradigm for the self-supervised enhancement of stimulus-related brain response data. Different time-aligned stimulus-evoked brain responses to the same stimulus are randomly shifted in time and independently enhanced. Both enhanced brain responses are compared using a model that predicts the shift in time between the brain responses. Using a model based on a multi-view convolutional neural network as an enhancement module, we show the efficacy of our method for a downstream task of decoding the speech envelope from auditory EEG. A significant relative improvement of 32% (p<0.001) was found when using the enhanced EEG versus normal EEG. While the shown example concerns EEG in response to auditory stimulation, conceptually, our method applies to other modalities (such as MEG) and other tasks (such as visual stimulus-response modelling).